Graph RAG

When to use

When relationships matter more than raw text: compliance, research, multi-hop questions.

Architecture sketch

Text → NER/RE → Triples → Graph DB
Query (entities/paths) → Retrieve → Compose → Evals

Pitfalls

• Noisy extraction → brittle paths
• Over-indexing without clear query policies

Checklist

• [ ] Define entity/edge schema
• [ ] Extraction evals on representative corpus
• [ ] Hybrid retrieve: graph paths + text

Good–Better–Best

• Good: simple entity linking
• Better: typed relations and path constraints
• Best: learned policies + feedback loops

Tiny eval snippet

def path_coverage(query, graph_paths):
    return any(required_edge in p for p in graph_paths)